Optimizing Parametric Model Selection in Anti-Cancer Drug Survival Analysis

OBJECTIVES: To evaluate the goodness-of-fit (GoF) of various survival models, including direct GoF for single treatments and relative GoF between regimens, compare the effectiveness of Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), and evaluate the impact of minimal number-at-risk considerations (consideration-minnrisk)

METHODS: GoF was measured by differences in Restricted Mean Survival Time (RMST), and Mean Absolute Error (MAE) between the parametrically extrapolated curves and updated Kaplan-Meier curves. Individual patient data were reconstructed from survival curves using established methods. Six model types evaluated included Standard Parametric Models (SPM), Fractional Polynomials (FP), Restricted Cubic Splines (RCS), Royston-Parmar models (RPM), Generalized Additive Models (GAM), and Parametric Mixture Models (PMM). Three GoF assessments (within-sample [GoF-f], extrapolation [GoF-e], and combined [GoF-fe]) were evaluated. Spearman's correlation assessed the relationships between GoF-f and GoF-e/GoF-fe. GoF were compared using Wilcoxon Signed-Rank test. Subgroup analyses and linear regressions assessed result robustness and identified influential variables. Finally, nomograms were constructed as internal verification and to predict optimal extrapolation models.

RESULTS: 99 trials providing 314 survival curves were included. No correlation was observed between GoF-f and GoF-e/GoF-fe across six models. Selecting AIC/BIC did not impact GoF-e/GoF-fe. Significant improvement in GoF-e/GoF-fe was shown for no-consideration-minnrisk. For GoF-f, GAM performed best while PMM performed worst. For direct GoF, extrapolation-MAE (MAE_e) rankings were: RP, SPM (p=0.01 [vs RP]), PMM (p<0.001), FP (p<0.001), GAM (p<0.001), RCS (p<0.001). For relative GoF, MAE_e rankings were: RP, SPM (p=0.24 [vs RP]), SPM+RP+RCS (p=0.006), SPM+RP (p=0.049), FP (p<0.001), SPM+FP+RCS+RP+GAM+PMM (p<0.001), RCS (p<0.001), GAM (p<0.001), PMM (p<0.001). Results for GoF-fe were consistent, as were as relative GoF, and subgroup findings. Regression results confirmed RPM's superiority. A dynamic nomogram was constructed (https://mingyezhaocpu.shinyapps.io/Dynamic_nomogram/), with good discrimination and calibration.

CONCLUSIONS: Less-utilized RPM shows superior GoF-e/GoF-fe, using machine learning for model selection may be promising and requires further study.